Lab 6: Telling a story about survival on the Titanic
Nicholas G Reich, for PUBHLTH 490ST at UMass-Amherst
Introduction to the Titanic dataset
In this exercise, we are going to look at a dataset describing characteristics (including survival) of passengers
on the Titanic, which sunk in the North Atlantic Ocean in 1912. You can hand in this lab for extra credit.
Getting started
Let’s load the data into our current R session, and look at the variables available in the dataset:
library("Hmisc")
getHdata(titanic)
head(titanic)
Exercise 1: Understand your data
For this analysis, we are going to focus primarily on the impact of three predictor variables on survival:
economic status (pclass), sex, and age. Examine your data carefully using some univariate plots and/or
summaries of the variables to understand what the distributions look like.
Exercise 1a: Deal with missing data
There are a lot of missing data in the age variable. In real data analysis problems, missing data is a common
and pesky problem. It can especially be difficult to deal with when the missingness is not “random”, i.e. certain
factors (whether they are variables you measure or not) can predict whether the data will be missing or
not. For example, do you think that older or younger people might be more or less likely to be missing their
age from this dataset? What other variables might determine whether we have age or not? One way to
start to look at this is to create a new factor variable that indicates whether age is missing or not for each
observation. Then we can create some simple tables to assess missingness across different groups. Try these
types of tabulations out and determine whether you think missing age is predicable based on some of the
other data in our dataset.
titanic$age_mis <- factor(is.na(titanic$age))
mosaic::tally(~sex|age_mis, data=titanic)
For now, to make the rest of the lab easier to work through, we are going to ignore the observations that
are missing age. This is rarely a great assumption to make in practice, especially if the missingness may be
associated with other factors. So when we interpret our results, we will need to remember that our dataset
may no longer be representative of the entire population of travelers on the Titanic. Run the following code
to remove anyone missing age from our dataset.
titanic1 <- dplyr::filter(titanic, age_mis==FALSE)
Exercise 2: Make some hypotheses
Before making any multivariate plots, discuss with your teammates your hypotheses about what relationships
might exist between these variables. Make a short list, including directions of possible relationships and
possible interactions. Sketch out a few graphics that you want to make.
1
Exercise 3: Look at your data
Create a few exploratory graphics and/or tables that illustrate the relationships between these variables and
survivial. (Hint: try adapting graphing code from the logistic regression lecture. Try using facets or colors to
highlight important comparisons.)
Exercise 4: Design and fit a model
Using the plots as your guide, write down a model that you’d like to fit to use to describe how this data
predicts the outcome of survival. Then fit that model.
Exercise 5: Examine your model performance
Now that you’ve fit a model, you can calculate, for each individual, an estimated probability of survival,
using code similar to that below (Note: adding type='response' ensures that if you fit a logistic regression
model, the predict function returns you predicted probabilities instead of predicted log-odds.):
titanic1$preds <- predict(fm1, type = "response")
Using these predicted probabilities, try to determine where your model performs worst. Start by figuring out
a metric that you could use to measure a poor prediction for a particular observation. Then, summarize your
data in a way that shows you what type of individuals you make the worst predictions about. Why did you
make poor predictions for these subsets of people? What other data would you like to have to help you make
even better predictions?
2